Please use this identifier to cite or link to this item:
|Title:||Colour retinal image segmentation for computer-aided fundus diagnosis|
|Keywords:||Hong Kong Polytechnic University -- Dissertations|
Diagnostic imaging -- Digital techniques.
Eye -- Imaging
Diabetic retinopathy -- Imaging -- Data processing.
|Publisher:||The Hong Kong Polytechnic University|
|Abstract:||Colour images of the ocular fundus, or retinal images, captured using digital fundus cameras reveal to us the retinal, ophthalmic, systemic diseases such as diabetes, hypertension, and arteriosclerosis and provide a non-intrusive way to screen retinopathy. Automated segmentation of colour retinal images can help ophthalmologists, oculists, or eye-care specialist to screen larger populations. The meaningful objects to be segmented include the main regions of the retina and the lesions caused by certain diseases. The appearance of certain lesions can be the sign of certain retinal diseases and systemic diseases. The main regions of the retina are the optic disc, fovea, and blood vessels. The identification of these regions can help in the analysis of diseases that affect these regions preferentially, such as glaucoma and proliferative diabetic retinopathy. The locations of these regions can then in turn also help in locating other lesions. The work in this thesis is in two parts. The first part addresses the segmentation of blood vessels, which are critical diagnostic features. The second part of this thesis proposes a system for segmenting the main regions and lesions of colour retinal images obtained from patients with diabetic retinopathy (DR). Retinal vessel segmentation: Automated retinal segmentation is difficult due to the fact that the width of retinal vessels can vary from very large to very small, and that the local intensity contrast of vessels can be weak and unstable. In this thesis, we will present a simple but efficient multiscale scheme, Multiscale Production of the Matched Filter (MPMF), that uses responses as the multiscale data fusion strategy. The proposed MPMF vessel extraction scheme includes: (1) multiscale matched filtering and scale multiplication in the image enhancement step; and (2) double thresholding in the vessel classification step. The fact that vessel structures have stronger responses to the matched filters at different scales than background noise means that multiplying the responses of matched filters at several scales enhances vessels while suppressing noise. Another difficulty of vessel segmentation is from the affection of lesions. For example, if we need to find the dark lines in an image, the edges of bright blobs will be the major source of false line detection. Consequently, some blobs (bright lesions and the optic disc) in the retinal image may cause false detection of vessels. In this thesis, we propose a modified matched filter to suppress the false detection caused by bright blobs. The proposed modified matched filter does not respond to non-line edges and so significantly reduces the false detection of vessels. Diabetic Retinopathy (DR) image segmentation: The objects useful for DR diagnosis include retinal lesions such as red lesions (intraretinal hemorrhages, microaneurysms), bright lesions (hard exudates and cottonwool spots) and retinal main regions such as vessels, optic disc, and fovea. Colour retinal image segmentation to assist DR diagnosis has attracted many researchers these years. But few works have been designed to extract all of these objects in one efficient scheme. The major disadvantages of current colour retinal image segmentation works are (1) they do not take into account the correlation among different objects and this leads to many false positives; (2) the algorithms are too time-consuming so that the online application is impossible. In this thesis, we propose one efficient scheme that segments all useful objects for DR diagnosis. Our segmentation scheme answers these issues by organizing the segmentation all objects in one efficient workflow. This scheme suppresses false positives effectively and improves segmentation speed. The segmentation speed is further improved by algorithm optimization and by keeping the algorithm as simple as possible.|
|Description:||xviii, 126 p. : ill. (some col.) ; 30 cm.|
PolyU Library Call No.: [THS] LG51 .H577P COMP 2010 Li
|Rights:||All rights reserved.|
|Appears in Collections:||Thesis|
Show full item record
Files in This Item:
|b23517360_link.htm||For PolyU Users||162 B||HTML||View/Open|
|b23517360_ir.pdf||For All Users (Non-printable)||5.68 MB||Adobe PDF||View/Open|
Checked on Feb 19, 2017
Checked on Feb 19, 2017
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.